SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient Descent

Analysis

Why This Paper Matters

LLM agents are increasingly used for recurring tasks, and the ability to accumulate procedural knowledge in the form of skills is crucial for efficiency and adaptability. Existing skill refinement methods rely on iterative execution and failure diagnosis but lack explicit outcome feedback and treat deletion as a generic edit. SkillProx addresses these gaps by introducing a proximal-gradient-inspired framework that couples closed-loop diagnostic evolution with utility-aware proximal refinement. This is significant because it provides a more principled and effective approach to skill learning, potentially improving agent performance in both in-distribution and out-of-distribution settings.

The paper's motivation is grounded in a composite objective that balances task loss and skill complexity, which is a novel perspective in the context of skill learning. By framing skill updates as a proximal gradient descent problem, SkillProx offers a theoretical foundation that is often missing in heuristic-based methods. This could inspire further research into optimization-inspired techniques for agent skill management.

Technical Contributions

• Forward-Backward Framework: SkillProx introduces a two-stage process that mirrors proximal gradient descent. The forward stage applies diagnosis-driven edits and re-executes them on the same task batch, rolling back regressions. The backward stage decomposes the skill into auditable knowledge units and estimates their contributions.
• Closed-Loop Diagnostic Evolution: Unlike prior methods, SkillProx explicitly uses outcome feedback from re-execution to inform subsequent diagnoses, creating a closed loop that improves edit quality.
• Utility-Aware Proximal Refinement: The backward stage uses a frozen leave-one-out utility audit to evaluate each knowledge unit's contribution, then applies validation-gated consolidation, demotion, or removal. This is a dedicated mechanism for deletion, which is more principled than treating it as a generic edit.
• Composite Objective: The method is motivated by a composite objective balancing task loss and skill complexity, providing a clear optimization target.

Results

SkillProx was evaluated on in-distribution and out-of-distribution benchmarks across multiple backbone LLMs. The main result is an average accuracy improvement of 3.0 percentage points over the strongest gradient-based baseline. Component ablations demonstrate that both closed-loop diagnosis and proximal refinement contribute complementary benefits, indicating that each component is essential for the overall performance gain. The consistency across different backbones and benchmarks suggests robustness of the approach.

Significance

SkillProx represents a step forward in making LLM agents more self-sufficient and adaptive. By providing a principled framework for skill evolution, it reduces the need for manual intervention and enables agents to continuously improve their procedural knowledge. The proximal-gradient inspiration could bridge the gap between optimization theory and agent skill learning, opening new avenues for research. The explicit handling of skill complexity and utility-based deletion also addresses practical concerns about skill bloat and relevance. As agents are deployed in more dynamic and long-horizon tasks, methods like SkillProx will be crucial for maintaining performance and efficiency.